615_final_project
Qixuan Zhang
12/16/2018
Abstract
This study investigates key factors that influence Airbnb customer experiences by analyzing big dataset of online review comments through the process of text mining and sentiment analysis. Based on the key attributes, we will give some suggestions to future Airbnb hosts. Compared with traditional study, the innovation of this study is using AFINN dataset to calculate sentiments scores of Airbnb customers’ comments to study customers’ opinions about Airbnb services. From our analysis, we found price is not the crucial factor influence customers’ experience, and customers’ overall review scores are based on location, accuracy, communication, room types and cleanliness. Methodologically, this study contributes to improve the satisfaction of customers and illustrate how big data can be used and visually interpreted in marketing studies. Key Words: Customers’ Reviews, Latent Rating Regression, Multilevel Regression
## ── Attaching packages ────────────────────────────────── tidyverse 1.2.1 ──## ✔ tibble 1.4.2 ✔ purrr 0.2.5
## ✔ tidyr 0.8.2 ✔ dplyr 0.7.8
## ✔ readr 1.1.1 ✔ stringr 1.3.1
## ✔ tibble 1.4.2 ✔ forcats 0.3.0## ── Conflicts ───────────────────────────────────── tidyverse_conflicts() ──
## ✖ dplyr::combine() masks gridExtra::combine()
## ✖ dplyr::filter() masks stats::filter()
## ✖ dplyr::lag() masks stats::lag()## Loading required package: NLP##
## Attaching package: 'NLP'## The following object is masked from 'package:ggplot2':
##
## annotate##
## Attaching package: 'scales'## The following object is masked from 'package:purrr':
##
## discard## The following object is masked from 'package:readr':
##
## col_factor## Loading required package: RColorBrewer## Type 'citation("pROC")' for a citation.##
## Attaching package: 'pROC'## The following objects are masked from 'package:stats':
##
## cov, smooth, var##
## Attaching package: 'MASS'## The following object is masked from 'package:dplyr':
##
## select## Loading required package: Matrix##
## Attaching package: 'Matrix'## The following object is masked from 'package:tidyr':
##
## expand## Loading required package: lme4##
## arm (Version 1.10-1, built: 2018-4-12)## Working directory is /Users/qixuanzhang/Desktop/615_final_project##
## Attaching package: 'arm'## The following object is masked from 'package:scales':
##
## rescale## Loading required package: sp## rgdal: version: 1.3-6, (SVN revision 773)
## Geospatial Data Abstraction Library extensions to R successfully loaded
## Loaded GDAL runtime: GDAL 2.1.3, released 2017/20/01
## Path to GDAL shared files: /Library/Frameworks/R.framework/Versions/3.5/Resources/library/rgdal/gdal
## GDAL binary built with GEOS: FALSE
## Loaded PROJ.4 runtime: Rel. 4.9.3, 15 August 2016, [PJ_VERSION: 493]
## Path to PROJ.4 shared files: /Library/Frameworks/R.framework/Versions/3.5/Resources/library/rgdal/proj
## Linking to sp version: 1.3-1## Loading required package: carData##
## Attaching package: 'car'## The following object is masked from 'package:arm':
##
## logit## The following object is masked from 'package:dplyr':
##
## recode## The following object is masked from 'package:purrr':
##
## some## Loading required package: zoo##
## Attaching package: 'zoo'## The following objects are masked from 'package:base':
##
## as.Date, as.Date.numeric##
## Attaching package: 'magrittr'## The following object is masked from 'package:purrr':
##
## set_names## The following object is masked from 'package:tidyr':
##
## extract##
## Attaching package: 'data.table'## The following objects are masked from 'package:dplyr':
##
## between, first, last## The following object is masked from 'package:purrr':
##
## transpose1. Introduction
1.1. Background
With advantages of price, there is no doubt Airbnb is becoming the best choice of more and more people. Due to high volume of customers, it is crucial for hosts to provide a comfortable and high-quality room for customers. Based on those previous studies, most of researchers paid more attention to listing dataset, they analyzed customers experience by using reviews rate as response variable and took economy-based variables as predictors.
In this study, we used sentiment analysis get the mean sentiments scores of each apartment in the cleaned listing dataset. And then we will use sentiment scores as response to fit our models and compare the results with the results based on traditional response (overall review rating/value). More specifically, we will compare key factors of customer reviews from different models. And then we figure out the key attributes that influence customers reviews and give some suggestions to improve customer reviews rate. #2. Previous Work ##2.1. Cleaning Dataset and Summary of Variables DATA-SOURCE DESCRIPTION: The Boston dataset was released by Airbnb itself on October 8th, 2018 to show people that how Airbnb is really being used and is affecting their neighborhood. The original data set had 6041 rows and 96 variables (columns) in Boston listing dataset, Boston reviews data had 173818 rows and 6 columns. CLEANING DATASET: We kept the variables that are relevance to customers reviews in listing dataset. Furthermore, we deleted NA values and filtered numbers of reviews more than three. In this study, we focused on price interval of $0-$1000 rental Airbnb apartments in Boston. The cleaned listing dataset we used had 3611 rows and 23 variables. And for Boston reviews dataset, we only deleted NA comments and non-English comments.
2.2. Summary Variables
Summary Variables Although there are 100+ variables in the 2 datasets, only 21 are used for the analysis. We briefly describe such variables is given below:
3. Exploratory Data Analysis
3.1. EDA of Boston Customers Reviews
3.1.1. Distribution of Airbnb Customers’ Reviews in Boston
To give our reader a more observable visualization, we made a leaflet to describe the overall review scores distribution in different area in Boston at first. With stronger and stronger of the green color, the overall review scores are higher.
Following leaflet plot of Boston customers’ reviews, we made a histogram plot to describe reviews of Boston.## Warning: Ignoring unknown parameters: binwidth, bins, pad
From Figure 3.1, we could get conclusion that most people prefer to give positive feedback. That made me more curious about what caused people to give negative feedback. Back to the Boston listing dataset, we found that reviews scores rating related to several aspects, therefore we made several plots to visualize and analyze different aspects related to customers’ reviews. ###3.1.2. Variable Analysis There are several categorical variables, from cleansed Boston listing table, therefore we made several plots to illustrate different relationship between customers reviews scores and variables.
Initially, we want to figure out overall review scores distribution of different neighborhoods. We made a bar plot to illustrate counts of different location related to their customers reviews scores for location, in Figure 3.2 we could find the best location of Airbnb apartments is in Allston. Because more than 2000 apartments located in Allston was given 10 scores by customers.
Based on the inspiration of neighborhood related to customers’ reviews scores, we analyzed plots of property type, room type and price related to customers’ reviews. From Figure 3.3, we found that the most common and popular room type is Entire home/apt, and there are few Shared room in Airbnb. 
And we found similar conclusion when we observed Figure 3.4, we found the most popular property type is Apartment. As for price analysis of different customers’ overall review scores , we made the followting plot.
After analyzing the price distribution of different customers’ review scores, we would like to moveforward to review scores rating of different property type. 
Additionally, we analyze the evaluation of different room type of different neighbourhood in Boston. 
###3.1.3. Compared different Reviews Scores Rating in Boston Based on above distribution of customers’ overall review scores in Boston, we classified the overall review scores to three parts, good, normal and bad. As for good part, it included review scores rating from customers that more than 90. And for the bad part, it included overall review scores rating less than 70. And within the interval of 70-90, we classified it marked by normal. And we added evaluation to listing table to illustrate the different. After classified overall review scores rating, we made a plot to compare the distribution of top 11 best reviews and top 11 worse reviews in Boston.
##3.2. Text mining Customers’ Review ###3.2.1. Word Counts Analysis of Customers’ Comments Based on visualization plots of different variables, we want to figure out some meaning information that latent in customers’ reviews. We believe that these comments made Airbnb customers give positive and negative feedback. And we also think based on comments sentiments analysis, we could get more specific idea that made us to get close to what happened during the customers’ Airbnb trip. Initially, we want to see the description combined with overall review scores. We divided our rating scores rating to two groups, first one is high rating group. We filtered the overall review scores rating more than 90 as the high rating group. And then we filter the overall review scores rating less than 70 as low rating group. After that, we apply these following methods to each city.
We get the plot of Top frequency words of high and low customers’ scores rating in Boston as following (Figure3.7). From comments words counts of low customers’ rating, we found some common word like noise, uncomfortable, etc. Compared with comments of low rating words clouds plot, we can find lots of positive words easily like clean, perfect, easy, beautiful, super, highly, super in comments of high rating words plot. 
Furthermore, we would like to build two would clouds to illustrate the specific frequency words of these above two parts. In Figure 3.9 and Figure 3.10, we displayed word clouds of high scores rating comments and low scores rating comments.word cloud of overall review scores rating and high overall review scores rating
## Warning in brewer.pal(1, "Dark2"): minimal value for n is 3, returning requested palette with 3 different levels
## Warning in brewer.pal(1, "Dark2"): minimal value for n is 3, returning requested palette with 3 different levels
When we complete the part of checking frequency words and words cloud, we move forward to sentiments analysis of Airbnb customers’ comments and we want to figure out the latent factors that influence customers’ overall review scores rating.
3.2.2. Sentiment Analysis
There is no doubt that comments of customers reflect why they give the positive or negative feedback to hosts. However, Airbnb did not provide a review system which could evaluate customers’ comments. In this study, we calculated sentiment scores by using Airbnb customers’ comments words joined with AFINN sentiment dataset, and then we create a new column to named sentiment which recorded sentiments scores corresponding to unique host ID. More specifically, we want to confirm if overall review scores rating is perfectly reflecting customers’ reviews.
After adding sentiment scores column, we want to analyze several customer behavior-related variables combined with overall review scores. More specifically, we want to confirm if overall eview scores rating is perfectly reflecting customers’ reviews. After getting sentiment scores of every host, we made the distribution of comments in different room type, plot that shown in Figure 3.10, by using AFFIN Lexicon sentiment scores. 
From Figure 3.11, we could get the conclusion that the distribution of sentiment scores follow normal distribution and Apartments is most popular property type.
#4.Benford Analysis 
When we applied beford analysis to our dataset, we found the price of our datset almostly followed benford law, but overall customers’ review scores didn’t following benford law. Based on overall customers’ review scores plot of Digits Distribution, we get the conclusion that overall reveiws scores interval of customers is concentrated on 88-100. From summation difference plot, we find scores from 88 to 100 are the most frequecy scores in customers’ reviews. In other word, the overall customers’ reviews scores from 88-100 is more meaningful and frequency.Therefore, we can not get the conclusion that this dataset is fraud. Because the lowest scores of customers’ reviews is 60(when we summarized our dataset, we found the minimum customers’ reviews scores is 60).As you can see in Chi-Squared Difference plot, customers’ review scores does not following benford law. To campared benford law within different variables, we also applied benford analysis on price in the same dataset.From price benford analysis, we could find that the most frequency digits used in Boston’s Airbnb price is 17. From the price’s Summation Distribution by Digits plot, we get the coclusion that the number 1 appears as the leading significant digit about 30% of the time, while 9 appears as the leading significant digit less than 5% of the time. From the above conclusion, we believe the price of Airbnb in Boston almostly follow Benford Law. #5 Conclusion ##5.1 Conclusion of Key Attributes In this study, based on the inspire of latent rating regression model (Wang,2011), I created a unique Airbnbsentiments scores of Airbnb customers’ comments and reviews scores of location, cleanliness and description etc. aspects from Airbnb to get more accuracy results. Based on above analysis, we will choose Latent Linear Regression model to analyze key attributes of customers’ reviews and give some suggestions to improve future customers’ reviews for Airbnb hosts.
The conclusion for key attributes of Aribnb in Boston is that only the overall review scores of different aspects did not include all the enough and reliable information. As we can see, combined with sentiment scores, we could get more accuracy predict results.
The findings from the present study provide important theoretical and managerial implications for academicians and practitioners. Specifically, the implications may be beneficial to scholars who seek to evolve research in the areas of customer experience and rating behavior, and to practitioners (particularly in the lodging industry) when developing useful strategies to create and offer the ideal experiences that customers seek. ##5.2 Future Direction The first goal is to combine more dataset of major cities in United States and create an Airbnb sentiment words analysis dataset. When we want to calculate the overall customers’ review scores of specific trips, we will combine the scores the customer given with his/her comment’s sentiment scores. Based on this reliable information, we could build a more sufficient, reasonable and reliable customer reviews scores rating system. Both Airbnb and hosts in Airbnb could get feedback to improve their service from the scores given by customers.